"One cannot act upon that which has not been defined."

My weekend list has six things on it. Laundry done. Floors vacuumed and washed. Grass cut and yard edged. Bathroom cleaned. Dishwasher run and emptied. Groceries every couple of weeks.

Four of those six say two things where one would have fit.

Floors could have said floors cleaned. It says vacuumed and washed, which is two operations in a fixed order, and the reason it says two is that I once finished vacuuming and considered the floors done, and they were not, and there was no argument to have with myself about it because the word clean had never been pinned to anything. Grass could have said yard done. It says cut and edged, and the edging is the part that gets skipped, which is exactly why it is written down separately. Dishwasher could have said dishes. It says run and emptied, because a dishwasher full of clean dishes is a machine that cannot be used and a kitchen that looks finished.

Bathroom says cleaned. One word.

I have run that list for years without noticing that I built four of its six items one way and the fifth another way, and I noticed only when I went looking for a place in my own life where this rule was already operating before I had a name for it. The list is a set of definitions. It is the same list I would have written as instructions for somebody else, which is the tell. Every item I have ever had to hand off got specified, and the one item nobody else has ever done for me is the one still sitting there as an adjective.

Now give that list to a person who does not live here.

Laundry done is fine. Floors vacuumed and washed is fine. Grass cut and yard edged is fine, with a question about how short, which they will resolve by looking at the neighbors. Bathroom cleaned produces a result, and it is somebody else's result, assembled out of somebody else's idea of what a clean bathroom is, and whatever comes back is not wrong. There is nothing to measure it against. If the mirror is streaked and the baseboards are untouched and the toilet is spotless, that is one legitimate reading of the word. If the toilet is ignored and the floor is mopped, that is another. The instruction was satisfied both times.

Nothing about that is a failure of effort or attention. The person did the work. What they could not do is the thing I never gave them, which is the definition, and the reason they could not do it is that they had to act, and so they supplied one. That is the part people miss about undefined instructions. They do not stall the work. They relocate the decision, quietly, to whoever is holding the mop, and then the result gets attributed to that person's care rather than to my sentence.

Move the same structure into a building. A field is added to a form because somebody needs to know how urgent a request is. The field is called Priority. It has four values. Nobody writes down what any of them mean, because everybody in the room at the time already knows, and they do know, and the form works beautifully for eleven months. Then the volume doubles, three more people start filling it in, and Priority stops being information. It becomes a record of who filled in the form.

Stated as a rule, the claim is narrower than the sentence sounds and stronger than it looks.

The variables are a term and an act. Some word appears in an instruction, a form, a policy, a conversation, or a task list, and somebody has to do something because of it. The word is the variable. The doing is the act.

The quantity minimized is the number of private resolutions. Every time an undefined term reaches somebody who has to act, that person resolves it, in their own head, according to their own history, and the resolution is invisible to everyone including them. Ten people, ten resolutions, one word. That spread is the cost, and it is not visible in any output the system produces, because each individual output looks correct.

The failure signal is a result that nobody can call wrong and nobody expected. Not an error. Errors announce themselves. This is the finished work that satisfies the instruction and misses the intent, arriving with a person attached to it who did exactly what was asked.

The domain of validity is coordinated action. The rule holds wherever more than one person must act on the same term, or where one person must act on it at two different times and needs the second to match the first. Outside that domain the rule weakens sharply, and the chapter spends its back half on why.

The word cannot is doing something specific in the original. Not should not, and not will act badly. The claim is that the act on an undefined term is not the act you think it is. Something got done, and what got done was determined somewhere other than where you thought you determined it.

The idea has a precise ancestor, and it came out of physics rather than management.

Percy Bridgman won a Nobel Prize for work on materials under extreme pressure, and in 1927 he wrote a book about what physics had just learned about its own vocabulary. Relativity had done something disturbing. It had shown that terms physicists had used confidently for two centuries, simultaneity above all, did not mean one thing. They meant whatever the procedure for measuring them made them mean, and when the procedure changed, so did the concept, and nobody had noticed because the procedures had all agreed with each other at ordinary speeds.

His conclusion was blunt. We mean by any concept nothing more than a set of operations, and the concept is synonymous with the corresponding set of operations. Length is not a property that exists independently and gets approximated by rulers. The concept of length is fixed when the operations by which length is measured are fixed, and if you measure the distance to a star by parallax and the distance to a goal post by laying down a meter stick, you are not doing the same thing twice with different precision. You are using one word for two concepts.

Bridgman added a test that reads like an engineering requirement. If a specific question has meaning, it must be possible to find operations by which an answer may be given to it. A question with no procedure attached is not a hard question. It is a question that has not finished being asked.

Read my weekend list against that and the four items that say two things are operational definitions and the one that says cleaned is not. Vacuumed names an operation. Washed names an operation. Cleaned names a state, and a state is what you get when somebody has already performed operations you did not specify.

W. Edwards Deming spent fifty years pushing the same point into industry, and he supplied the structure that makes it usable by people who are not physicists. An operational definition, in his treatment, has three parts: a test that can be performed on the thing, a criterion for judging the result of that test, and a decision, yes or no. Not a description. A procedure with an outcome. He was making the point about words like safe, round, reliable, and clean, which are the words specifications are made of and which mean nothing at all until somebody says how they would be checked.

He went further than Bridgman in one direction that matters here. In his later work he held that there is no true value of any characteristic, state, or condition that is defined in terms of measurement or observation. The number you get is a joint product of the thing and the procedure. Change the procedure and you get a different number, and neither number is the real one, because there is no real one to be had. That sounds like relativism and is the opposite. It is what forces the procedure into the specification. If the value depends on the method, then publishing a value without the method publishes nothing.

There is an older claim in the neighborhood, the idea that a definition works by naming a category and then naming what distinguishes the thing from the rest of that category, and it is routinely credited to Aristotle. I went looking for the passage and could not verify the attribution in the form it is usually given. Aristotle used the vocabulary; the systematic theory appears to be Porphyry's, centuries later. So it is not in this chapter. A creed that claims to test its own foundations does not get to keep a citation because the citation is flattering.

The empirical case for defining things early is strong in one respect and much weaker than it is usually presented.

The strong part is that requirements defects are real, they are the most expensive class, and the cost of one rises with how long it goes unnoticed. A study out of NASA's Johnson Space Center examined error cost escalation across the life of aerospace projects using three separate methods, and every method produced the same shape: an error caught while the requirement is being written costs one unit, and the same error caught in operations costs many multiples of that.

The shape is the finding. The multiplier is where the honesty has to come in, because the commonly quoted version of this research has been smoothed into a rule of thumb that the research does not support. The three methods disagree wildly at the far end. One gives twenty-nine times. One gives between a hundred and fifty-seven and a hundred and eighty-six. One gives sixteen hundred and fifteen. Those are not converging estimates of a single quantity. They are three different quantities produced by three different accounting choices, and anyone quoting a single figure for what a late-caught requirements error costs is quoting a choice, not a measurement.

The authors say so themselves. They note that the data was limited in quantity and in the detail available about the errors involved, that it came from two large aerospace projects, and that the third method leans on the author's experience rather than on collected data. They recommend the work be repeated on smaller projects and in other industries. As far as I can tell it largely has not been.

Two further limits apply before any of this reaches a person's weekend list or a company's intake form. The evidence is about software and aerospace engineering, where a requirement is a written artifact that persists and can be traced to the defect it caused. Most undefined terms in ordinary work leave no artifact at all. Nobody can trace the cost of Priority meaning four things, because the cost shows up as a pile of requests that moved at the wrong speed, each of which has a perfectly good local explanation. The research measures the cases where the accounting is possible, which are not the typical cases.

And the direction of causation is assumed rather than shown. That requirements errors cost more the later they are found is well supported. That writing better definitions up front reduces total cost is a different claim, and it requires knowing what the definition work itself costs, which the studies do not report. The next chapter, the one about defining every variable to its deepest identifiable condition, is where that bill comes due.

The clearest version of this I have worked on is an intake form.

A service request arrives. Before anything can happen, the request has to be classified: what kind of problem it is, how urgent, whose it is, and whether it is covered. Those four answers determine everything downstream, including who is dispatched, what gets ordered, how fast it moves, and whether anyone is billed. The form asks for them.

The first attempt to write down what the categories meant produced an argument that lasted longer than building the form. Not about the categories. About one word inside one of them.

The word was covered. Everyone agreed the form should capture whether a request was covered. Everyone had been using the word daily for years. When the question became what test a person could run on a request to answer yes or no, it turned out that at least three different tests were in circulation, all of them reasonable. One person meant the installation was inside its warranty period. One meant the specific failure was of a type the warranty addresses, regardless of date. One meant the customer had a contract that would pay for the visit whether or not the warranty applied. Those three tests disagree on real requests, frequently, and the disagreements had been invisible because each person had been answering their own question correctly and nobody had ever had to write down which question it was.

That argument was the deliverable. The form was the easy part. What came out of it was three fields where there had been one, because there were three questions, and merging them into one word had been hiding a decision rather than simplifying it.

The same thing happens on my weekend list at a scale of one person. Bathroom cleaned hides a decision. It has never cost me anything, because I am the only one who has ever resolved it and I resolve it the same way each time without having to think. The cost appears the instant a second person is involved, and it appears as a bathroom that is clean by a definition I never wrote down and would not have chosen.

The boundary is the expert acting alone under time pressure, and it is real enough to take the sentence apart.

The rule holds because an undefined term reaching two people produces two resolutions. Remove the second person and the argument loses its engine. A person who has done something ten thousand times has a resolution of every term in the job, stable, consistent with itself, and unavailable for inspection. They cannot tell you what makes a weld sound. They can tell you that one is not.

The rule also assumes that definition precedes action, and there are whole categories of work where the sequence runs the other way. A diagnosis is a definition produced by acting. A negotiation defines its terms through the offers made in it. Research is the activity of acting on concepts specifically in order to find out what they are, which is exactly what Bridgman's operations were describing when physics discovered that simultaneity had been two concepts the whole time. Physics did not define its way to relativity. It acted, got a result it could not explain, and went back to the words afterward.

So the domain narrows to this: wherever a term must produce the same act in more than one person, or in one person across time, it has to be defined before it is acted on. Where the actor is one expert, operating inside their own history, on a problem they have seen, the definition exists and is simply not in words. That is a real difference and not a loophole, and the next section is where somebody who has spent his career studying those experts gets to make the case that I have the whole thing pointed the wrong way.

Before that, three things the rule cannot see.

It cannot tell a term that needs defining from one that does not, and the cost of guessing wrong runs in both directions. Carl Hempel made the sharpest version of this objection against Bridgman's operationalism directly: taken strictly, it produces a proliferation of concepts of length, of temperature, and of every other scientific concept, that would be not only practically unmanageable but theoretically endless. Every new procedure is a new concept. Measure length with a ruler and with light and you have two lengths. Define clean for the bathroom, then for the kitchen, then for the truck, and the definitions do not compose. Pushed to its end, the rule dissolves the vocabulary it was meant to sharpen. This is the collision with my own ism about defining every variable to its deepest identifiable condition, and it is the reason that one needs a cost curve rather than a direction.

It cannot define itself without circularity. Somebody has to decide what counts as defined, and that decision is made by a person acting on an undefined term. In practice the resolution is Deming's three parts, and it is a procedure rather than a philosophy: a term is defined when a test exists, a criterion exists, and the answer comes out yes or no. That works. It is also plainly a stopping rule adopted for convenience, and I would rather say that than pretend the regress has a bottom.

And it says nothing about definitions that are wrong. A term can be perfectly operationalized and perfectly useless, or perfectly operationalized around the wrong quantity, and the rule is satisfied either way. Everyone will agree on what was measured and everyone will be measuring the thing that does not matter. Defining a term is how you stop an argument about what the word means. It is not how you find out whether the word was worth using.

Here is the case against.

Gary Klein, Roberta Calderwood and Anne Clinton-Cirocco went to fireground commanders in the mid-1980s and asked how they made decisions on the way to and inside a burning building. The researchers expected what the decision literature of the time expected: that a commander confronting a serious fire would generate options, compare them, and select. They studied twenty-six commanders with an average of twenty-three years of experience each, and they probed a hundred and fifty-six separate decision points drawn from real incidents.

In fewer than twelve percent of those decision points was there any evidence of two or more options being compared at all. In over eighty percent, the commander recognized the situation as an instance of something they had seen, and recognized an action as the one that goes with it, and did it.

Klein's model of what is happening, which he has spent thirty years refining, is not that these people are guessing. It is that recognition retrieves a course of action, and then a mental simulation runs it forward to check whether it will work, and if it survives that check it is executed without any second option ever being generated. The evaluation is real. The comparison is what is missing.

Now put the sentence to that commander. One cannot act upon that which has not been defined. He is standing in front of a structure that is doing something he would call wrong, and he cannot tell you what wrong means, and he has not defined unsafe or about to flash over or sound enough to send people into, and he is acting, correctly, at a speed no definitional process could reach. He does not merely act on undefined terms. Acting on undefined terms is the whole capability. It is what twenty-three years bought him.

And the objection has a second edge that cuts deeper than the first. Force him to define. Put the terms in a manual, require that classification precede action, insist that unsafe resolve to a checklist. Two things happen, and both are documented in the literature on expertise. The performance gets slower, which on a fireground is itself a failure mode. And the checklist captures a fraction of what the recognition was using, because most of what it was using was never in propositional form and cannot be put there. You have replaced a rich instrument with a poor one and called the substitution rigor. Anyone who has watched an organization write procedures for its best people has watched this happen.

Definition is what you do when you have no expertise, and dressing it up as a universal principle mistakes a remedy for scarcity into a virtue. Where real skill exists, demanding definitions degrades it. One cannot act upon that which has not been defined is then not a law about action. It is a description of what novices and committees need, generalized into a creed by somebody who has spent his career building systems for organizations rather than standing in fires.

I think that lands. I also think it is answerable, and the answer costs me the sentence in its current form.

Klein's fireground commanders are not acting on undefined terms. They are acting on terms defined in a medium that does not use words.

Recognition is a definition. It has a test, which is whether the pattern matches; a criterion, which is the accumulated record of what followed the last several thousand times it matched; and a decision, yes or no, which arrives in under a second. That is Deming's three-part structure exactly, implemented in tissue rather than in a document, built by twenty-three years of paying the tuition. The commander's unsafe is operationally defined to a standard most written specifications never reach. What it is not is transferable, inspectable, or available to anyone else.

Which locates the narrowing precisely. The original sentence claims that action requires definition, and that is too strong, because a definition can live in one person and work. What action across people requires is a definition that exists outside any one person's head. The commander can act. The commander cannot hand the act to the next shift, cannot defend it in a review, cannot be audited, and cannot be replaced. Those four things are what the written definition buys, and they are the things organizations run on and individuals mostly do not need.

That also settles which way one cannot act upon that which has not been defined should be applied, which I have been sloppy about for years. It is not an instruction to define everything before doing anything. It is a test to run at the moment work crosses from one person to another, or from one moment to a later one where consistency matters. My weekend list obeys the rule at exactly the four places where the work was ever going to be handed off, and ignores it at the one place where it never has been. I did not plan that. The list found the boundary on its own, which is some evidence the boundary is where I now claim it is.

Where Klein wins outright, and I will not argue it: inside the expertise, under time pressure, on a problem the expert has seen, forcing verbal definition makes the decision worse. The correct move there is to leave it alone, which is another of my own isms arriving early, the one holding that where a data-based decision will not produce a genuinely better result, you leave it manual. Chapter 15 is where that boundary gets sharp enough to use, because stated loosely it excuses anyone who would rather not write anything down.

What survives:

The original claimed too much and therefore said less than it meant. The revision gives up the universal and keeps the part that does work: the moment a term has to produce the same act in a second person, or in the same person at a later time, an undefined term is not a small imprecision. It is a decision that has been made somewhere nobody is looking.

Two questions leave this chapter open. The regress has no bottom and I have adopted a stopping rule instead of solving it, which is honest but not finished. And nothing here tells you how deep a definition should go, only that it should exist. Chapter 3 takes depth, and with it the bill.